Deep Learning-Based Optimization Framework for Full-Waveform Inversion in Tunnels

Recently, geophysical exploration methods have been prevalently used in ahead geological prospecting in tunneling. Among all kinds of geophysical exploration methods, the seismic method is more popular because of its interface sensitivity and long detection distance. However, because of the tunnel observation limitation and the detection data lacking, the seismic method represented by the full-waveform inversion (FWI) does not work well with satisfactory precision for tunnel geology detection and cannot meet tunnel construction safety requirements. Therefore, studying the high-precision seismic method suitable for tunnel environments is in great demand. To address these problems, we propose optimizing the full-waveform inversion by deep learning from the inversion interface and velocity value. To optimize the interface, we built a virtual tunnel observation system similar to the observation system on the ground with a large offset distance to obtain accurate interface gradients. Then, we learn the mapping between the actual tunnel gradients and virtual gradients. Finally, the virtual gradients are used to update the velocity model. To optimize the velocity value, we adopt another specific network to help full-wave inversion get out of local minima. Extensive model and field tests, as well as interpretability studies, verify the feasibility and effectiveness of the proposed method.